How to catch spec conflicts before they become $50K rework

On a recent 3,200 sq ft remodel in Boston, the architectural plan showed a 9-foot finished ceiling on the second floor. The structural set, drawn by a different firm two months later, showed a 14-inch flush LVL header running the full length of the same span. The mechanical drawings, drawn by a third firm a month after that, showed an 8-inch round duct routed exactly through that header. Nobody caught it during plan review. The GC noticed the conflict three days after framing was up.

The fix cost $32,400 in rework, $11,000 in re-engineering, four days of crew time, and a permit revision. A 90-second cross-sheet comparison would have flagged it at submittal. That conflict, and the four others on the same job, were not exotic. They were the normal output of three separate firms working from slightly different basis-of-design assumptions, none of them wrong on their own, all of them disastrous together.

Why manual review misses these

The standard plan review for a residential project is one person — usually the PM, sometimes the GC — going through the architectural set page by page, then flipping to structural to spot-check obvious load paths, then to mechanical for the same. The review covers maybe 60% of the dimensional information in each set, and almost none of the implicit conflicts between sets. A few specific failure modes:

• Header depth vs ceiling height.Architect spec's a finished ceiling that assumes a 12-inch joist depth. Structural spec's a 16-inch beam in the same span. The 4-inch difference becomes a 2x4 soffit drop nobody quoted.
• Duct routing vs structural members.MEP routes a 10-inch supply trunk through a space where the structural drawing has a 7.25-inch flush beam. Both are right on their own sheet. Together they don't fit.
• Plumbing chase vs framing.A 4-inch waste stack needs to go through a wall the structural set shows as a 2x4 shear wall with no holes permitted. The fix is a thicker wall and another sheet of plywood you didn't budget for.
• Door swing vs cabinet layout.Architectural shows a 36-inch door swinging into a kitchen. Interior elevations show the refrigerator at the hinge side of that door. The fridge doesn't open more than 60 degrees. Caught at appliance install, not at framing.
• Window header height vs trim detail. Architectural specifies 6-foot 8-inch head height at all windows. Trim detail specifies a 7-inch crown that needs to die into the casing. The casing now sits 1 inch above the door head. Looks wrong at walk-through, gets re-trimmed.

What manual review is actually good at

Manual plan review by an experienced GC is excellent at three things: spotting buildability issues that the design team didn't consider, catching code violations in their region, and predicting which subs are going to flag the spec. It is bad at numerical cross-referencing across hundreds of pages of dimensions.

That second category is what AI cross-sheet analysis is for. It is not a replacement for the experienced eye. It is the experienced eye plus a tireless reviewer that holds all 350 pages of the drawing set in working memory simultaneously and flags every number that does not reconcile.

How AI cross-sheet analysis works

The mechanism is straightforward. The plan set is OCRed and chunked. Every dimensional callout, every material spec, every schedule entry becomes a structured fact. The system then runs a set of cross-sheet checks:

1. Spatial reconciliation.For every room or assembly, the system compares the architectural finished dimensions against the structural framing dimensions and the mechanical routing requirements. Anything that doesn't reconcile within a configurable tolerance gets flagged.
2. Schedule consistency.Door schedule, window schedule, finish schedule, and the plans are cross-checked. A door tagged D-12 on the plan with a 3-0 width should appear on the schedule with a 3-0 width. When it doesn't, the system asks.
3. Code-related dimensional rules. Egress windows with sill heights, stair rises, headroom at landings — anywhere a code rule has a hard dimensional constraint, the system can check against the drawing.
4. Citation back to the source page.Every flagged conflict cites the specific page number, sheet number, and location on the sheet for each side of the conflict. The reviewer can verify the AI's finding in 30 seconds rather than 30 minutes.

What it gets right, and what it doesn't

Be clear-eyed about this. A well-tuned AI plan reader will catch 75-85% of cross-sheet dimensional conflicts on a clean architectural set. On a messy set — handwritten markup, low-resolution scans, revision clouds layered three deep — the catch rate drops to maybe 50%. Even at 50%, you are still catching three to five conflicts that a human reviewer was going to miss, on a typical residential project.

One catch pays for the year. Three catches pay for the year ten times over. That is not marketing math — that is the math you want to run against your own rework line for the last twelve months.

What it does not do, and you should not pretend it does:

• It will not catch design intent issues that aren't dimensional. If the architect specified an unusable kitchen layout, AI won't flag it because it is geometrically valid.
• It will not catch issues that depend on local code interpretation from your inspector. AI knows the IRC. It does not know your inspector's preferences.
• It will not catch issues that are only visible in 3D — a duct that physically can't make a turn in the space allowed. The 2D dimensional check is most of the value, but it is not all of it.

How to roll this into your workflow

The right time to run a cross-sheet review is after permit issuance but before lumber order. That window is typically 1-3 weeks on a residential project. The flow looks like:

1. Upload the permitted set to your plan reader the day permits issue.
2. Run the cross-sheet conflict pass. Expect 5-15 flagged items on a typical 3,000-4,000 sq ft project.
3. Review each flagged item with your PM. Triage into "real conflict, fix at design," "field adjustment OK," and "false positive." A practiced reviewer can clear 10 flags an hour.
4. For real conflicts, submit RFIs to the architect or structural engineer in writing. Reference the page numbers the system cited.
5. Order materials after the RFIs are resolved. The cost of waiting two days for an answer is dramatically less than the cost of ordering wrong.

How Buildra fits

Buildra's plan reader is purpose-built for this workflow on residential projects. It chunks the architectural, structural, and MEP sets, runs the cross-sheet reconciliation, and surfaces conflicts with page citations and one-click RFI generation. Most users catch their first $20K+ conflict within two weeks of starting their first project on the platform.